YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Vision-OPD Intrinsic-Reward Experiments — Best Checkpoints (ready to eval)
Merged HF-format weights (bf16 safetensors + tokenizer + image processor configs) of the best checkpoint of each of the 6 experiments: {TTRL, Intuitor, RENT} × {Qwen3.5-4B, Qwen3.5-9B} trained on Vision-OPD-6K. Best = highest 6-benchmark mean on the held-out selection split (3216 items, judge Qwen2.5-72B-Instruct). Each folder loads directly with transformers / vLLM.
| folder | method | model | recipe | best step | avg6 (select) |
|---|---|---|---|---|---|
ttrl-4b-nokl-best-s20/ |
TTRL | Qwen3.5-4B | no KL | 20 | 74.46 |
intuitor-4b-nokl-best-s5/ |
Intuitor | Qwen3.5-4B | no KL | 5 | 70.25 |
rent-4b-kl-best-s5/ |
RENT | Qwen3.5-4B | KL β=0.01 | 5 | 71.03 |
ttrl-9b-nokl-best-s15/ |
TTRL | Qwen3.5-9B | no KL | 15 | 78.46 |
intuitor-9b-kl-best-s5/ |
Intuitor | Qwen3.5-9B | KL β=0.01 | 5 | 76.85 |
rent-9b-kl-best-s5/ |
RENT | Qwen3.5-9B | KL β=0.01 | 5 | 77.87 |
Benchmarks: V*Bench, ZoomBench, HR-Bench 4K, HR-Bench 8K, MME-RealWorld, MME-RealWorld-CN.
Usage
from huggingface_hub import snapshot_download
path = snapshot_download("q1716523669/visionopd-best-ckpts",
allow_patterns="ttrl-9b-nokl-best-s15/*")
# then point vLLM / transformers at f"{path}/ttrl-9b-nokl-best-s15"
Eval settings used in our record: vLLM, temperature 0.0, seed 0, max_tokens 2048, max-model-len ≥ 32768 (HR-Bench 8K inputs reach >16k tokens).
Full 13-checkpoint FSDP archive (incl. optimizer state), logs and eval outputs: q1716523669/visionopd-archive.